Abstracts

The centenial lowest interest rates and the amplitude of the most rapid descent since 15 september 2008 , as well as their crawling behaviour since 5 years seem to have created a surprise. The Bank Risk Manager aims to rely on fundamentals of interest modelling like « Economie et Intérêt, Allais, 1947 », the « John Taylor’s rule », 1993. The data mining of long-term time-series was performed on LIBOR and swap interest rates with daily data for a dozen of maturities with 20 years history for some major currencies : CHF ; USD ; Euro (10 Y), GBP, Yen, and with yearly or other frequencies for Swiss Francs mortgage rates and other rates since 1850. Findings of such data mining include the incontestable downward wavy trends since 20 years from now, which fit into the 160 years record of Swiss mortgage rates, showing now the third local long term minimum, in addition to those of 1890 and 1950. The date-maturity interest rates maps and their statistics depict the populations which vary according to maturity, from three different populations for LIBOR to one population for long term swaps. The bizarre moves on GBP can reveal the recent fraud on LIBOR. Adjustment to asymetric thick tails distributions could be achieved. In accordance with Bachelier’s finding, the daily yields (or moves) do not show time-memory with the autocorrelation computations, this is the contrary for the memory of the I.R. values themselves. The statistical autocorrelation of the signs and of the absolute daily moves reveal a one-week memory of the absolute values of change. The series of consecutive draw-up and draw-downs behave differently as medium terme memory was depicted by the statistics on the number of consecutive bullish or bearish days. Herealso the absolute daily moves show an autocorrelation structure. SOM Self-Organizing Maps maps confirm the clustering of time-periods , and ANN Artificial Neural Networks achieve an adapted fitting of trends and disturbancies, the MLP MultiLayer Perceptron stick on long-term trends during temptative forecasts. As a conclusion, Data Mining Methods with spatial statistics extract the usefull information for model development, out of the half-million figures analyzed.